Hugging Face is GitHub plus Kaggle plus a model zoo plus a social network — all in the AI era.
Models: 1M+ models hosted, from Qwen3 to gpt-oss to quantized fine-tunes of everything. The weights download in minutes. The README quality varies wildly — read the model card carefully.
Datasets: the real treasure. Open datasets for instruction tuning, code, math, multilingual text, and tool-calling. This is where the research value lives.
Spaces: running demos you can fork and modify. Best way to try a new model without installing anything — then you clone it and run it yourself.
Inference Providers: HF now routes to serious providers (Together, Fireworks, Groq, Azure) from one token. They've stopped being just a registry and started being infrastructure.
The underrated corner: HF Hub discussions and model cards as signal. Before betting a project on a model, read the issues. You'll learn whether the author responds to bug reports, whether the license is actually permissive, and whether the quantized GGUF is sane.
Verdict: if you do any open-weights work, HF is unavoidable. Treat it as a library, not a marketplace — verify before you trust.